Prediction of particle mixing process in a rotating drum based on convolutional neural network

Abstract

The Discrete Element Method (DEM) has been widely used to analyze particle mixing processes. However, in chemical industry applications, mixing often involves billions of particles, making DEM simulations computationally expensive due to the intensive processes of contact detection and force calculation. Convolutional Neural Network (CNN) leverages images from DEM simulations as input, maintaining computational efficiency regardless of particle number, thus offering an effective solution to reduce computational costs. To balance accuracy and efficiency, we propose a Convolutional Neural Network with a Multi-Branch Block (CNNMB), which integrates convolutional kernels of different sizes through skip connections to extract global features. The performance of CNNMB was evaluated using three key metrics. Results show that across 24 DEM test cases, the predicted mixing index achieved an R² greater than 0.998, the predicted dynamic angle of repose reached an accuracy exceeding 98.4 %, and the predicted average particle height yielded an R² above 0.96. Furthermore, to capture the time-sequential characteristics of the mixing index, we developed a hybrid architecture by coupling CNNMB with a Long Short-Term Memory (LSTM) network—referred to as the Convolutional Neural Memory Network (CNMN). Results indicate that CNMN achieved a prediction accuracy of over 82 % across simulation cases with varying parameters. Additionally, a large-scale simulation involving one million particles was conducted, demonstrating that CNMN reduced computational time by approximately 97-fold compared to DEM simulations, highlighting its potential for efficient predicting.